Home
last modified time | relevance | path

Searched refs:opt_elems (Results 1 – 2 of 2) sorted by relevance

/libCEED/backends/cuda-shared/
H A Dceed-cuda-shared-basis.c73 const CeedInt opt_elems[7] = {0, 32, 8, 6, 4, 2, 8}; in CeedBasisApplyTensorCore_Cuda_shared() local
75 … CeedInt elems_per_block = CeedIntMax(thread_1d < 7 ? opt_elems[thread_1d] / num_comp : 1, 1); in CeedBasisApplyTensorCore_Cuda_shared()
127 const CeedInt opt_elems[7] = {0, 32, 8, 6, 4, 2, 8}; in CeedBasisApplyTensorCore_Cuda_shared() local
129 … CeedInt elems_per_block = CeedIntMax(thread_1d < 7 ? opt_elems[thread_1d] / num_comp : 1, 1); in CeedBasisApplyTensorCore_Cuda_shared()
165 const CeedInt opt_elems = block_size / (Q_1d * Q_1d); in CeedBasisApplyTensorCore_Cuda_shared() local
166 const CeedInt elems_per_block = opt_elems > 0 ? opt_elems : 1; in CeedBasisApplyTensorCore_Cuda_shared()
171 const CeedInt opt_elems = block_size / (Q_1d * Q_1d); in CeedBasisApplyTensorCore_Cuda_shared() local
172 const CeedInt elems_per_block = opt_elems > 0 ? opt_elems : 1; in CeedBasisApplyTensorCore_Cuda_shared()
338 const CeedInt opt_elems[7] = {0, 32, 8, 6, 4, 2, 8}; in CeedBasisApplyAtPointsCore_Cuda_shared() local
340 … CeedInt elems_per_block = CeedIntMax(thread_1d < 7 ? opt_elems[thread_1d] / num_comp : 1, 1); in CeedBasisApplyAtPointsCore_Cuda_shared()
[all …]
/libCEED/backends/hip-shared/
H A Dceed-hip-shared-basis.c225 const CeedInt opt_elems = block_size / Q_1d; in CeedBasisApplyTensorCore_Hip_shared() local
226 const CeedInt elems_per_block = opt_elems > 0 ? opt_elems : 1; in CeedBasisApplyTensorCore_Hip_shared()
231 const CeedInt opt_elems = block_size / (Q_1d * Q_1d); in CeedBasisApplyTensorCore_Hip_shared() local
232 const CeedInt elems_per_block = opt_elems > 0 ? opt_elems : 1; in CeedBasisApplyTensorCore_Hip_shared()
237 const CeedInt opt_elems = block_size / (Q_1d * Q_1d); in CeedBasisApplyTensorCore_Hip_shared() local
238 const CeedInt elems_per_block = opt_elems > 0 ? opt_elems : 1; in CeedBasisApplyTensorCore_Hip_shared()